MIT: By 2029, AI will be capable of handling most text-based tasks.

The latest research from MIT shows that AI's automation substitution in the labor market is not a sudden "giant wave," but rather a widespread and continuous "rising tide." It is expected that by 2029, the AI success rate for most text-based work tasks will reach 80% to 95%.

Research Background: How Will AI Change Work?

In March 2026, MIT’s FutureTech team released a working paper titled “Crashing Waves or Rising Tides: Early Findings on AI Automation from Thousands of Workers' Labor Market Task Assessments” (Paper Number: arXiv:2604.01363v1), with authors including Matthias Mertens, Neil Thompson, etc. Research funding came from Open Philanthropy and a tech company.

This study attempts to answer a question crucial to both investors and policymakers: Will improvements in AI capabilities bring about a “crashing waves” model—suddenly and intensely disrupting certain jobs—or a “rising tides” model—broadly and gradually raising the overall level of automation?

The two models affect the labor market in very different ways. “Crashing waves” means certain professions go from hardly affected to almost completely replaced in the short term, leaving workers little time to adapt; “rising tides” makes changes more predictable but covers a broader range, with equally profound final impacts.

Figure 1 “Crashing Waves” VS “Rising Tides”

Research Methods: 17,000 Real Worker Assessments

The research team selected 11,768 text-based tasks from the O*NET database’s 18,786 tasks with at least 10% time-saving potential. Eventually, 11,536 tasks were included, generating 69,216 task instances.

The assessment process was strict: Each task instance was answered by 41 large language models (LLMs), then scored by real practitioners with relevant work experience, using a 1 to 9 scale—7 or above means “usable with minimal editing,” 9 means “better than ordinary human workers.” The study included over 17,000 valid assessments, with about 34.6% of raw data excluded for quality reasons.

Task durations ranged from about 10 minutes to several days, with most tasks taking between 20 minutes and 10 hours.

The research team emphasized these are preliminary results, with data collection still ongoing. The current sample is slightly lower than the target distribution in wage level and educational requirements, with white-collar jobs overrepresented.

Figure 3 Task Duration Histogram

Core Finding 1: It’s “Rising Tide,” Not “Giant Wave” 

The core finding is: The curve of AI success rate versus task duration is surprisingly flat.

Specifically, for every tenfold increase in task duration, the log-odds of AI success rate (at ≥7 points threshold) drops by only 0.31. At a 60% average success rate, this predicts an acceptance rate drop of only about 7.6 percentage points. This curve is much flatter than those of previous benchmark-based studies, such as METR.

The paper states: “For a large number of representative, realistic labor market tasks that LLMs can handle, the decline trend between task success rate and task duration is on average surprisingly gentle—in other words, it’s more like water rising than waves crashing.”

This pattern holds true across models of different sizes and publication dates. Most job categories (such as management, community and social services) also follow the “rising tide” pattern.

However, there are significant differences between job categories. The highest success rates are “Installation, Maintenance, and Repair” (72.5%) and “Construction & Extraction” (71.0%), lowest is “Legal” (46.8%). The steepest slope is “Personal Care and Service” (β=-0.93), meaning task duration impacts AI success rate most in that field.

Figure 4 Task Instance Automation by Required Task Completion Time

Core Finding 2: Progress Is Faster Than Expected, Task Processing Duration “Doubles” Every 3.8 Months

“Rising tide” does not mean slow. The data shows AI capabilities are improving quickly.

The study writes: “During Q2 2024 to Q3 2025, frontier models raised the success rate for 3–4 hour tasks from 50% to 1 week tasks, and raised the success rate for 1 minute tasks from 70% to 1 hour tasks.”

• At the 50% success rate threshold, task duration handled by frontier models jumped from 3–4 hours to 1 week
• At the 70% success rate threshold, task duration jumped from 1 minute to 1 hour

If you take the “human task duration” that AI can process at a certain success rate (e.g., 50%) as a metric, the “doubling time” is only 3.8 months. This is among the faster speeds in current studies—METR previously reported doubling times of 2–6 months, Kwa et al. reported 4–7 months.

From the perspective of failure rate, failure rate halving time for tasks from 5 minutes to 24 hours is 2.4 to 3.2 years, corresponding to a success rate increase of about 8–11 percentage points per year.

This broad and rapid improvement confirms the “rising tide” logic: New model releases produce a parallel upward shift of the success rate curve.

Figure 6 Task Duration and Success Rate Thresholds over Time

Core Finding 3: Large Models and New Models Have Different Paths of Improvement

The study also distinguishes two paths of capability improvement, which has direct significance for AI investment logic.

Model Size (Large vs. Small): Large models with over 100 billion parameters are clearly superior to small models for short-term tasks, but their advantage narrows for long-term tasks—the curve “rotates outward” (β=-0.36 vs -0.26).

Model Age (New vs. Old): Models released after 2025 show roughly equal gains across task durations compared to older models—the curves shift upwards almost in parallel.

This means iteration over time (newer models) more evenly improves AI performance on longer complex tasks than expansion in size (larger models). This finding is notable for investors concerned with AI infrastructure ROI.

Prediction for 2029: Gradual, but Not to be Underestimated

Extrapolating from current trends, the research team made a clear time prediction:

It is estimated that by 2029, the AI success rate for most tasks will reach 80%-95%, meeting the minimum qualified quality level. Most tasks in our survey take hours, so by 2029 the success rate will be nearly 90%.

However, the study also states it will take years more to achieve “near perfect” success rate (close to 100%). This provides workers a window period to adjust, especially in fields with low tolerance for error.

The team notes these forecasts assume the pace of AI progress stays at the near-two-year trend, so they should be viewed as upper-bound scenarios. Potential slowdowns include rising compute costs, slower hardware gains, reduced algorithm innovation, and physical limits on chip performance.

Figure 7 Predicted AI Success Rates Over Time

Impact on the Labor Market: Task Automation ≠ Worker Replacement

The research team emphasizes that high AI success rates at the task level cannot be directly equated to the same proportion of jobs being automated.

There are three reasons:

1. Data bias: Current samples may over-represent readily surveyed professions, which may also be easier to automate, leading to potentially inflated estimates
2. “Last mile” costs: Integrating AI into real workflows involves extra costs such as information acquisition, system integration, compliance, and sometimes it is not economically viable
3. Difference between tasks and professions: Automating a single task does not necessarily reduce employment in an entire occupation. As shown by Autor and Thompson (2025), the impact of task automation on wages and employment depends on that task’s position within the occupational task bundle, which can lead to wage rises or drops, job increases or decreases

The paper notes: “It is still unclear how the economy will respond, but it is hard to imagine continued AI progress aligning stably with today’s economic status quo.”

Risk Warning and DisclaimerThe market has risks, investment requires caution. This article does not constitute personal investment advice, nor does it consider the individual user’s special investment goals, financial status, or needs. Users should consider whether any opinions, viewpoints, or conclusions in this article suit their own specific conditions. Proceed at your own risk.